Product Recommendation Based on Embeddings: People Who Viewed This Product Also Viewed These Products
Ulukbek Attokurov, Okan Kaya, Mehmet Selman Sezgin · 2022
Nowadays, recommendation systems are getting more attention from both academia and industry. However, most industrial recommendation systems are still using non-embedding models such as TF-IDF-based machine learning models. We also have had an existing model implemented using TF-IDF. This current model served as the baseline for our performance evaluations. In this paper, we implemented an embedding-based “People Who Viewed This Product Also Viewed These Products” recommendation. The purpose of this recommendation is to be able to recommend alternative products to customers that are similar to the product they are viewing. In this way, users will not only be able to see their own search and filtering results but also view the alternatives based on the experience of the users who view similar products. The product view sequences to be used in model training are created using 6 months of product view data. The number of daily product view events is over 20 million. We used offline and online metrics to compare the new model results with the existing model. We have achieved a significant lift in product coverage with 286%, order count with 29%, and distinct ordered product count with 30% without a decrease in CTR value.